"build, then rebuild 3 times in less time than it would have taken to build manually" - wish this is explored more. I saw the typical guy one-shot code generation (seen one with undefined variable in happy path) and throw it over the wall. I'm trying to use my 1200 words per minute typing gift for harder tests, readable code, refactors, scripts etc. Starting with my side project though.

Meh, they work, sort of, sometimes.

Don't get me wrong. I've been using Claude Code and Codex CLI for quite some times now and it's amazing what they can sometimes do. (I skipped the "Copilot" phase where AI was just a "better" auto-complete)

Emphasis on sometimes. And you really have to double check everything they do. So much.

And literally this week, Claude turned "dumb". Things I'd expect it to be able to do before, result in stuff I really just throw out the window. I thought I maybe started prompting differently or something so I tried multiple times on the same task. But no, it just went nowhere this week. Codex worked fine on the same problem but it tried to cheat real bad on the test cases. Luckily I caught it but otherwise the tests would've been completely useless. Essentially "always green".

And this is on the "pay-per-token" work account, so I can't simply explain it away with "they're saving on compute for free / bulk pay".

It will be like automatic looms. People doing basic repetative things with mediocre and below skills and refusing to upskill are in trouble. People who are creative and constantly drive their skillset upwards i think willě safe and sound.

I oscillate between worry and an overwhelming sense of job security.

[dead]